add pruned-24 NVFP4 AWQ encoder

NOTICE

@@ -4,6 +4,6 @@ Upstream sources:
 - Qwen/Qwen3-VL-8B-Instruct, revision 0c351dd01ed87e9c1b53cbc748cba10e6187ff3b
 - Qwen/Qwen3-VL-32B-Instruct, revision 0cfaf48183f594c314753d30a4c4974bc75f3ccb
 
-Modifications include language-model extraction, physical layer pruning, INT8 ConvRot conversion, trained ARA residual weights, and a trained 4096-to-5120 conditioning adapter. These modifications are not produced or endorsed by the Qwen team.
+Modifications include language-model extraction, physical layer pruning, NVFP4/AWQ and INT8 ConvRot packaging, trained ARA residual weights, and a trained 4096-to-5120 conditioning adapter. These modifications are not produced or endorsed by the Qwen team.
 
 MiniMax-H3 is the compatible downstream model. No MiniMax-H3 diffusion-model or VAE weights are included.

README.md

@@ -9,6 +9,7 @@ tags:
 - text-encoder
 - qwen3-vl
 - int8
+- nvfp4
 - convrot
 - comfyui
 ---
@@ -31,6 +32,7 @@ text_encoders/
 │   └── minimax_h3_recovered_8b_manifest.json
 └── pruned24_balanced/
     ├── qwen3vl_32b_minimax_h3_pruned24_balanced_int8_convrot.safetensors
+    ├── qwen3vl_32b_minimax_h3_pruned24_balanced_nvfp4_awq.safetensors
     ├── balanced_ara.safetensors
     └── minimax_h3_pruned24_balanced_manifest.json
 ```
@@ -62,9 +64,26 @@ Download these from [Comfy-Org/MiniMax-H3](https://huggingface.co/Comfy-Org/Mini
 | Recovered 8B BF16 | Qwen3-VL-8B, 24 language layers | ARA + adapter | 5120 | Text-only T2V |
 | Recovered 8B INT8 ConvRot | Same 24 layers, native Comfy INT8 ConvRot | ARA + adapter | 5120 | Text-only T2V |
 | Pruned-24 balanced | Qwen3-VL-32B-derived sparse 24-block INT8 ConvRot | balanced ARA | 5120 | Text-only T2V |
+| Pruned-24 balanced NVFP4/AWQ | Same retained blocks `0–8,35–49`, pruned directly from the official NVFP4/AWQ checkpoint | balanced ARA | 5120 | Text-only T2V |
 
 `ConvRot` describes base-weight storage and runtime quantization. It does not replace the recovered 8B adapter. The adapter is a nonlinear 4096-to-5120 module and is executed by the loader.
 
+### Pruned-24 NVFP4/AWQ memory validation
+
+The old pruned-24 INT8 ConvRot artifact was physically pruned from the full 50-layer INT8 ConvRot checkpoint (`bc2ced0…`), not from the official NVFP4/AWQ checkpoint. Because each retained INT8 layer is larger than an NVFP4 layer, that artifact did not realize the expected VRAM reduction.
+
+The NVFP4/AWQ variant is instead a sparse copy of the pinned official `Comfy-Org/MiniMax-H3` checkpoint at revision `0543966fbdce5ba05709a8f2031c94bdba629b4a`. It preserves all retained quantized tensors exactly and removes all tensors for original language blocks `9–34`.
+
+Clean external VRAM sampling used one fresh process per encoder, the same seven-token prompt, and 100 ms `nvidia-smi` sampling:
+
+| Encoder | Incremental peak GPU memory |
+|---|---:|
+| Official Qwen3-VL-32B NVFP4/AWQ | `16,430 MiB` |
+| Pruned-24 INT8 ConvRot + balanced ARA | `15,525 MiB` |
+| Pruned-24 NVFP4/AWQ + balanced ARA | `10,150 MiB` |
+
+The new NVFP4/AWQ variant produced `[1, 7, 5120]` conditioning and had cosine similarity `0.9999949932098389` against the pruned INT8 ConvRot variant for that prompt. These are short-prompt technical measurements, not visual-quality claims. See `evidence/pruned24_nvfp4_validation.json`.
+
 ## Validation and limits
 
 The showcase used one matched 1344x768, 243-frame, 24 FPS, 25-step T2V contract. The three source outputs and merged comparison are technical assets. They do not establish a visual-quality winner. The recovered 8B showcase uses the BF16 base; the INT8 ConvRot variant has separate numerical conditioning-parity evidence.

SHA256SUMS

@@ -1,24 +1,26 @@
 9841c64dada1d47f0ff8260af284c5db33808c59d4b0fe3c0e3a1a46fa04a4a0  .gitattributes
 c71d239df91726fc519c6eb72d318ec65820627232b2f796219e87dcf35d0ab4  LICENSE
-99a47c9f0361c7ccabaf7664ef428f825ffa3bd74dfbb7add813b310438882d7  NOTICE
-4ed0b3302f048d3e83c9cc639b95ed040961aa94a78f7ea186620c6460828e7b  README.md
+d71a9933d42aed06eb4dd813250c2ff13d424510339b19759723fede95a5e93b  NOTICE
+9416ea0dd0176ac310b346c9d1e36ced8fceb2ff63c93950b8eebb7a93968b69  README.md
 4c24aedf5abe9dbbb384aca8e1f9a8366b40080ee9e0961d425d05827ea8d940  evidence/comparison_validation.json
 ebf35ad3b14d354f868dfed7e7fe28a2e4bd5c26f4f14c62b9b3e44f46bb203f  evidence/cut_diagnostics.json
 649b6a8cacf1f254279b6c9087465f568554a708f0ff2e33cc23a0dab26a8bc9  evidence/final_manifest.json
 f7f7bc9b683ac65aa2eb13a270e55a8a584f3f4985b7e9ad649c3b002bb10b25  evidence/final_private_release_smoke.json
+d3c4f9053fa68e527270832262741d5dc31aac34d2ff17c251ebef22b6c0497c  evidence/pruned24_nvfp4_validation.json
 fcda410f971b45a674b023802069046b9384aa63c40f0412ec89616845141ffb  evidence/recovered_8b_int8_convrot_build_manifest.json
-c3802513601b4d596da1d95f4b89b7a2009210dcd3a1268179c4e025fd7e7d78  evidence/release_loader_smoke.json
+4aa3d707baccd850d141ef23f568b7ef2dcc9dd8665ae3efa75c0e3b9882124c  evidence/release_loader_smoke.json
 7625f40f4e11dfda1f7f01e532adb65a043616b55fe21a904c57f90e54fc0dd9  evidence/technical_validation.json
 63dc23f2939827bc6c601a8fbbfad8e64f9220554b20f349748d736cdb7e17c6  evidence/three_encoder_teacher_cosine.json
-1a024cdd580b92c50c22fed02d15644aab58760bb80d41df4c9b764e531fbbbb  model_index.json
+6ec16fa128fb0dd47e0a46617237f09757a1a981e489202dfaf61600c32b0b55  model_index.json
 01d731b4280c7d35b70057f2294ad4df6051ecb1215db3407fe7eca2cecc07e6  showcase/images/three_text_encoders_20_shots_each.png
 9acc6f48853f7def4a13bc2b537486a250c0ada0641427f2ca3afab80084ecd6  showcase/videos/minimax_h3_text_encoder_comparison_reference_recovered8b_pruned24_30s.mp4
 e42558dd5a7bf5b28d385c3851a4da13169881b1ea646a8c9161882c25c8053a  showcase/videos/official_qwen3vl_32b_nvfp4_awq_10s.mp4
 8091a8ab44519e7629d5e36c26940764f98d3abf0a0b2c7937f81a940a094124  showcase/videos/pruned24_balanced_ara_int8_convrot_10s.mp4
 1ab0cbc71335418921c99a74191f50d77f5aa58f3a5d6a035266758ada34020f  showcase/videos/recovered_qwen3vl_8b_bf16_ara_adapter_10s.mp4
 38afb86416aacafcb952e64fc481f698ecb90a61111319f73e83cfa5c589ebbd  text_encoders/pruned24_balanced/balanced_ara.safetensors
-b15dcb10102f2868ab7ee1fbf2241ecd7bcd30114feb23ccb66a648c55cac48c  text_encoders/pruned24_balanced/minimax_h3_pruned24_balanced_manifest.json
+31a520790540e83488872212f36fede04eec8325dc8ed3686938c05267649fed  text_encoders/pruned24_balanced/minimax_h3_pruned24_balanced_manifest.json
 cbafb2213438e16c82356a25057b3feafb926bf7f09c4db96c3f58a248450744  text_encoders/pruned24_balanced/qwen3vl_32b_minimax_h3_pruned24_balanced_int8_convrot.safetensors
+9e3c36932e10849fd2a946bb45d3e22834c95485d412a9293d1fcd082eb698e9  text_encoders/pruned24_balanced/qwen3vl_32b_minimax_h3_pruned24_balanced_nvfp4_awq.safetensors
 4e31eac82f99f38e68e65d2f1352dba3b307c06c6a5b5bb03890c422a419d863  text_encoders/recovered_8b/ara.safetensors
 59167808bca712ecd2078997f234d90ccfe46f42715e9e63fc6def4867655866  text_encoders/recovered_8b/conditioning_adapter.safetensors
 acef4566b75b4973d2f790bfac09f18a9c872146a4aa14e606c4e6ac488a4150  text_encoders/recovered_8b/minimax_h3_recovered_8b_manifest.json

evidence/pruned24_nvfp4_validation.json

@@ -0,0 +1,54 @@
+{
+  "format": "minimax_h3_te_clean_external_vram_v1",
+  "method": "Fresh child process per variant; nvidia-smi sampled every 100 ms; identical seven-token prompt; baseline subtracted.",
+  "variants": [
+    {
+      "id": "official_nvfp4_awq",
+      "baseline_mib": 502,
+      "peak_mib": 16932,
+      "incremental_peak_mib": 16430,
+      "exit_code": 0
+    },
+    {
+      "id": "pruned24_int8_convrot",
+      "baseline_mib": 503,
+      "peak_mib": 16028,
+      "incremental_peak_mib": 15525,
+      "exit_code": 0
+    },
+    {
+      "id": "pruned24_nvfp4_awq",
+      "baseline_mib": 503,
+      "peak_mib": 10653,
+      "incremental_peak_mib": 10150,
+      "exit_code": 0
+    }
+  ],
+  "limitations": [
+    "External process-wide GPU memory, not tensor residency attribution.",
+    "Short text-conditioning smoke; longer multimodal conditioning can require more memory.",
+    "Current local ComfyUI/PyTorch/comfy-kitchen stack on NVIDIA GeForce RTX 5090 Laptop GPU."
+  ],
+  "artifact": {
+    "file": "text_encoders/pruned24_balanced/qwen3vl_32b_minimax_h3_pruned24_balanced_nvfp4_awq.safetensors",
+    "bytes": 8553755869,
+    "sha256": "9e3c36932e10849fd2a946bb45d3e22834c95485d412a9293d1fcd082eb698e9",
+    "source_sha256": "35a88d51044231fe332301d7a62aa81e3f2cba62febeb446e2c1e3e0ef76f2c6",
+    "source_revision": "0543966fbdce5ba05709a8f2031c94bdba629b4a",
+    "source_keys": 2054,
+    "output_keys": 1170,
+    "removed_keys": 884,
+    "retained_tensor_mismatches": 0,
+    "nvfp4_linear_count": 168
+  },
+  "conditioning": {
+    "shape": [
+      1,
+      7,
+      5120
+    ],
+    "pruned_int8_vs_nvfp4_cosine": 0.9999949932098389,
+    "mean_abs_difference": 0.3806621730327606,
+    "max_abs_difference": 56.3076171875
+  }
+}

evidence/release_loader_smoke.json

@@ -1,6 +1,7 @@
 {
   "choices": [
     "pruned24_balanced: int8_convrot",
+    "pruned24_balanced: nvfp4_awq",
     "recovered_8b: recovered_8b_bf16",
     "recovered_8b: recovered_8b_int8_convrot"
   ],
@@ -22,6 +23,18 @@
         7
       ]
     },
+    "pruned24_balanced: nvfp4_awq": {
+      "conditioning_dtype": "torch.float32",
+      "conditioning_shape": [
+        1,
+        7,
+        5120
+      ],
+      "mixed_quantization": true,
+      "token_tags_shape": [
+        7
+      ]
+    },
     "recovered_8b: recovered_8b_bf16": {
       "conditioning_dtype": "torch.float32",
       "conditioning_shape": [

model_index.json

@@ -15,6 +15,11 @@
       "id": "pruned24_balanced_int8_convrot",
       "manifest": "text_encoders/pruned24_balanced/minimax_h3_pruned24_balanced_manifest.json",
       "base_file": "text_encoders/pruned24_balanced/qwen3vl_32b_minimax_h3_pruned24_balanced_int8_convrot.safetensors"
+    },
+    {
+      "id": "pruned24_balanced_nvfp4_awq",
+      "manifest": "text_encoders/pruned24_balanced/minimax_h3_pruned24_balanced_manifest.json",
+      "base_file": "text_encoders/pruned24_balanced/qwen3vl_32b_minimax_h3_pruned24_balanced_nvfp4_awq.safetensors"
     }
   ],
   "required_custom_node": "https://github.com/kgonia/ComfyUI-MiniMaxH3TextEncoders",

release_manifest.json

@@ -11,14 +11,14 @@
       "sha256": "c71d239df91726fc519c6eb72d318ec65820627232b2f796219e87dcf35d0ab4"
     },
     {
-      "bytes": 606,
+      "bytes": 619,
       "path": "NOTICE",
-      "sha256": "99a47c9f0361c7ccabaf7664ef428f825ffa3bd74dfbb7add813b310438882d7"
+      "sha256": "d71a9933d42aed06eb4dd813250c2ff13d424510339b19759723fede95a5e93b"
     },
     {
-      "bytes": 5143,
+      "bytes": 6640,
       "path": "README.md",
-      "sha256": "4ed0b3302f048d3e83c9cc639b95ed040961aa94a78f7ea186620c6460828e7b"
+      "sha256": "9416ea0dd0176ac310b346c9d1e36ced8fceb2ff63c93950b8eebb7a93968b69"
     },
     {
       "bytes": 6874,
@@ -40,15 +40,20 @@
       "path": "evidence/final_private_release_smoke.json",
       "sha256": "f7f7bc9b683ac65aa2eb13a270e55a8a584f3f4985b7e9ad649c3b002bb10b25"
     },
+    {
+      "bytes": 1723,
+      "path": "evidence/pruned24_nvfp4_validation.json",
+      "sha256": "d3c4f9053fa68e527270832262741d5dc31aac34d2ff17c251ebef22b6c0497c"
+    },
     {
       "bytes": 1786,
       "path": "evidence/recovered_8b_int8_convrot_build_manifest.json",
       "sha256": "fcda410f971b45a674b023802069046b9384aa63c40f0412ec89616845141ffb"
     },
     {
-      "bytes": 1088,
+      "bytes": 1368,
       "path": "evidence/release_loader_smoke.json",
-      "sha256": "c3802513601b4d596da1d95f4b89b7a2009210dcd3a1268179c4e025fd7e7d78"
+      "sha256": "4aa3d707baccd850d141ef23f568b7ef2dcc9dd8665ae3efa75c0e3b9882124c"
     },
     {
       "bytes": 5784,
@@ -61,9 +66,9 @@
       "sha256": "63dc23f2939827bc6c601a8fbbfad8e64f9220554b20f349748d736cdb7e17c6"
     },
     {
-      "bytes": 1104,
+      "bytes": 1372,
       "path": "model_index.json",
-      "sha256": "1a024cdd580b92c50c22fed02d15644aab58760bb80d41df4c9b764e531fbbbb"
+      "sha256": "6ec16fa128fb0dd47e0a46617237f09757a1a981e489202dfaf61600c32b0b55"
     },
     {
       "bytes": 5270610,
@@ -96,15 +101,20 @@
       "sha256": "38afb86416aacafcb952e64fc481f698ecb90a61111319f73e83cfa5c589ebbd"
     },
     {
-      "bytes": 1614,
+      "bytes": 2273,
       "path": "text_encoders/pruned24_balanced/minimax_h3_pruned24_balanced_manifest.json",
-      "sha256": "b15dcb10102f2868ab7ee1fbf2241ecd7bcd30114feb23ccb66a648c55cac48c"
+      "sha256": "31a520790540e83488872212f36fede04eec8325dc8ed3686938c05267649fed"
     },
     {
       "bytes": 14455970440,
       "path": "text_encoders/pruned24_balanced/qwen3vl_32b_minimax_h3_pruned24_balanced_int8_convrot.safetensors",
       "sha256": "cbafb2213438e16c82356a25057b3feafb926bf7f09c4db96c3f58a248450744"
     },
+    {
+      "bytes": 8553755869,
+      "path": "text_encoders/pruned24_balanced/qwen3vl_32b_minimax_h3_pruned24_balanced_nvfp4_awq.safetensors",
+      "sha256": "9e3c36932e10849fd2a946bb45d3e22834c95485d412a9293d1fcd082eb698e9"
+    },
     {
       "bytes": 29367232,
       "path": "text_encoders/recovered_8b/ara.safetensors",
@@ -140,15 +150,16 @@
     "text_encoders": true,
     "vaes": false
   },
-  "status": "private_release",
+  "status": "public_release",
   "validation": {
-    "extension_contract_tests": "3 passed",
+    "extension_contract_tests": "4 passed",
     "final_comfy_runtime_smoke": {
       "edge": "MiniMaxH3TextEncoderLoader.CLIP -> MiniMaxH3ImageToVideo.clip",
       "evidence": "evidence/final_private_release_smoke.json",
       "status": "passed",
       "variant": "recovered_8b: recovered_8b_int8_convrot"
     },
+    "pruned24_nvfp4": "evidence/pruned24_nvfp4_validation.json",
     "quantization_build_manifest": "evidence/recovered_8b_int8_convrot_build_manifest.json",
     "recovered_bf16_vs_int8_convrot": {
       "cosine": 0.9999650716781616,
@@ -169,6 +180,18 @@
           7
         ]
       },
+      "pruned24_balanced: nvfp4_awq": {
+        "conditioning_dtype": "torch.float32",
+        "conditioning_shape": [
+          1,
+          7,
+          5120
+        ],
+        "mixed_quantization": true,
+        "token_tags_shape": [
+          7
+        ]
+      },
       "recovered_8b: recovered_8b_bf16": {
         "conditioning_dtype": "torch.float32",
         "conditioning_shape": [

text_encoders/pruned24_balanced/minimax_h3_pruned24_balanced_manifest.json

@@ -8,10 +8,21 @@
     "license": "Apache-2.0",
     "modifications": ["language-model extraction", "physical 50-to-24-block pruning", "INT8 ConvRot conversion"]
   },
-  "checkpoint": {
-    "file": "qwen3vl_32b_minimax_h3_pruned24_balanced_int8_convrot.safetensors",
-    "sha256": "cbafb2213438e16c82356a25057b3feafb926bf7f09c4db96c3f58a248450744",
-    "format": "INT8 ConvRot"
+  "base_variants": {
+    "int8_convrot": {
+      "file": "qwen3vl_32b_minimax_h3_pruned24_balanced_int8_convrot.safetensors",
+      "sha256": "cbafb2213438e16c82356a25057b3feafb926bf7f09c4db96c3f58a248450744",
+      "format": "INT8 ConvRot",
+      "source_sha256": "bc2ced0fbea64757fa9acddccfc0b3f4819d1dcf1da6c124d690d368be283923"
+    },
+    "nvfp4_awq": {
+      "file": "qwen3vl_32b_minimax_h3_pruned24_balanced_nvfp4_awq.safetensors",
+      "sha256": "9e3c36932e10849fd2a946bb45d3e22834c95485d412a9293d1fcd082eb698e9",
+      "format": "NVFP4/AWQ",
+      "source_repo_id": "Comfy-Org/MiniMax-H3",
+      "source_revision": "0543966fbdce5ba05709a8f2031c94bdba629b4a",
+      "source_sha256": "35a88d51044231fe332301d7a62aa81e3f2cba62febeb446e2c1e3e0ef76f2c6"
+    }
   },
   "retained_blocks": [0, 1, 2, 3, 4, 5, 6, 7, 8, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49],
   "removed_blocks": [9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34],
@@ -28,6 +39,7 @@
   "limitations": [
     "Text-only T2V. Image and reference inputs are unsupported.",
     "Current package uses 50 execution shells with identity blocks for removed layers.",
+    "NVFP4/AWQ requires compatible NVIDIA NVFP4 hardware and a compatible ComfyUI/comfy-kitchen build.",
     "ARA remains a separate runtime residual. It is not merged into the INT8 ConvRot base."
   ]
 }

text_encoders/pruned24_balanced/qwen3vl_32b_minimax_h3_pruned24_balanced_nvfp4_awq.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9e3c36932e10849fd2a946bb45d3e22834c95485d412a9293d1fcd082eb698e9
+size 8553755869